4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains functions related to writing back dirty pages at the
9 * 10Apr2002 akpm@zip.com.au
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/spinlock.h>
18 #include <linux/swap.h>
19 #include <linux/slab.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/init.h>
23 #include <linux/backing-dev.h>
24 #include <linux/blkdev.h>
25 #include <linux/mpage.h>
26 #include <linux/percpu.h>
27 #include <linux/notifier.h>
28 #include <linux/smp.h>
29 #include <linux/sysctl.h>
30 #include <linux/cpu.h>
31 #include <linux/syscalls.h>
34 * The maximum number of pages to writeout in a single bdflush/kupdate
35 * operation. We do this so we don't hold I_LOCK against an inode for
36 * enormous amounts of time, which would block a userspace task which has
37 * been forced to throttle against that inode. Also, the code reevaluates
38 * the dirty each time it has written this many pages.
40 #define MAX_WRITEBACK_PAGES 1024
43 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
44 * will look to see if it needs to force writeback or throttling.
46 static long ratelimit_pages
= 32;
48 static long total_pages
; /* The total number of pages in the machine. */
49 static int dirty_exceeded
; /* Dirty mem may be over limit */
52 * When balance_dirty_pages decides that the caller needs to perform some
53 * non-background writeback, this is how many pages it will attempt to write.
54 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
55 * large amounts of I/O are submitted.
57 static inline long sync_writeback_pages(void)
59 return ratelimit_pages
+ ratelimit_pages
/ 2;
62 /* The following parameters are exported via /proc/sys/vm */
65 * Start background writeback (via pdflush) at this percentage
67 int dirty_background_ratio
= 10;
70 * The generator of dirty data starts writeback at this percentage
72 int vm_dirty_ratio
= 40;
75 * The interval between `kupdate'-style writebacks, in centiseconds
76 * (hundredths of a second)
78 int dirty_writeback_centisecs
= 5 * 100;
81 * The longest number of centiseconds for which data is allowed to remain dirty
83 int dirty_expire_centisecs
= 30 * 100;
86 * Flag that makes the machine dump writes/reads and block dirtyings.
91 * Flag that puts the machine in "laptop mode".
95 EXPORT_SYMBOL(laptop_mode
);
97 /* End of sysctl-exported parameters */
100 static void background_writeout(unsigned long _min_pages
);
102 struct writeback_state
104 unsigned long nr_dirty
;
105 unsigned long nr_unstable
;
106 unsigned long nr_mapped
;
107 unsigned long nr_writeback
;
110 static void get_writeback_state(struct writeback_state
*wbs
)
112 wbs
->nr_dirty
= read_page_state(nr_dirty
);
113 wbs
->nr_unstable
= read_page_state(nr_unstable
);
114 wbs
->nr_mapped
= read_page_state(nr_mapped
);
115 wbs
->nr_writeback
= read_page_state(nr_writeback
);
119 * Work out the current dirty-memory clamping and background writeout
122 * The main aim here is to lower them aggressively if there is a lot of mapped
123 * memory around. To avoid stressing page reclaim with lots of unreclaimable
124 * pages. It is better to clamp down on writers than to start swapping, and
125 * performing lots of scanning.
127 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
129 * We don't permit the clamping level to fall below 5% - that is getting rather
132 * We make sure that the background writeout level is below the adjusted
136 get_dirty_limits(struct writeback_state
*wbs
, long *pbackground
, long *pdirty
,
137 struct address_space
*mapping
)
139 int background_ratio
; /* Percentages */
144 unsigned long available_memory
= total_pages
;
145 struct task_struct
*tsk
;
147 get_writeback_state(wbs
);
149 #ifdef CONFIG_HIGHMEM
151 * If this mapping can only allocate from low memory,
152 * we exclude high memory from our count.
154 if (mapping
&& !(mapping_gfp_mask(mapping
) & __GFP_HIGHMEM
))
155 available_memory
-= totalhigh_pages
;
159 unmapped_ratio
= 100 - (wbs
->nr_mapped
* 100) / total_pages
;
161 dirty_ratio
= vm_dirty_ratio
;
162 if (dirty_ratio
> unmapped_ratio
/ 2)
163 dirty_ratio
= unmapped_ratio
/ 2;
168 background_ratio
= dirty_background_ratio
;
169 if (background_ratio
>= dirty_ratio
)
170 background_ratio
= dirty_ratio
/ 2;
172 background
= (background_ratio
* available_memory
) / 100;
173 dirty
= (dirty_ratio
* available_memory
) / 100;
175 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
176 background
+= background
/ 4;
179 *pbackground
= background
;
184 * balance_dirty_pages() must be called by processes which are generating dirty
185 * data. It looks at the number of dirty pages in the machine and will force
186 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
187 * If we're over `background_thresh' then pdflush is woken to perform some
190 static void balance_dirty_pages(struct address_space
*mapping
)
192 struct writeback_state wbs
;
194 long background_thresh
;
196 unsigned long pages_written
= 0;
197 unsigned long write_chunk
= sync_writeback_pages();
199 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
202 struct writeback_control wbc
= {
204 .sync_mode
= WB_SYNC_NONE
,
205 .older_than_this
= NULL
,
206 .nr_to_write
= write_chunk
,
209 get_dirty_limits(&wbs
, &background_thresh
,
210 &dirty_thresh
, mapping
);
211 nr_reclaimable
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
212 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
)
217 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
218 * Unstable writes are a feature of certain networked
219 * filesystems (i.e. NFS) in which data may have been
220 * written to the server's write cache, but has not yet
221 * been flushed to permanent storage.
223 if (nr_reclaimable
) {
224 writeback_inodes(&wbc
);
225 get_dirty_limits(&wbs
, &background_thresh
,
226 &dirty_thresh
, mapping
);
227 nr_reclaimable
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
228 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
)
230 pages_written
+= write_chunk
- wbc
.nr_to_write
;
231 if (pages_written
>= write_chunk
)
232 break; /* We've done our duty */
234 blk_congestion_wait(WRITE
, HZ
/10);
237 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
)
240 if (writeback_in_progress(bdi
))
241 return; /* pdflush is already working this queue */
244 * In laptop mode, we wait until hitting the higher threshold before
245 * starting background writeout, and then write out all the way down
246 * to the lower threshold. So slow writers cause minimal disk activity.
248 * In normal mode, we start background writeout at the lower
249 * background_thresh, to keep the amount of dirty memory low.
251 if ((laptop_mode
&& pages_written
) ||
252 (!laptop_mode
&& (nr_reclaimable
> background_thresh
)))
253 pdflush_operation(background_writeout
, 0);
257 * balance_dirty_pages_ratelimited - balance dirty memory state
258 * @mapping: address_space which was dirtied
260 * Processes which are dirtying memory should call in here once for each page
261 * which was newly dirtied. The function will periodically check the system's
262 * dirty state and will initiate writeback if needed.
264 * On really big machines, get_writeback_state is expensive, so try to avoid
265 * calling it too often (ratelimiting). But once we're over the dirty memory
266 * limit we decrease the ratelimiting by a lot, to prevent individual processes
267 * from overshooting the limit by (ratelimit_pages) each.
269 void balance_dirty_pages_ratelimited(struct address_space
*mapping
)
271 static DEFINE_PER_CPU(int, ratelimits
) = 0;
274 ratelimit
= ratelimit_pages
;
279 * Check the rate limiting. Also, we do not want to throttle real-time
280 * tasks in balance_dirty_pages(). Period.
282 if (get_cpu_var(ratelimits
)++ >= ratelimit
) {
283 __get_cpu_var(ratelimits
) = 0;
284 put_cpu_var(ratelimits
);
285 balance_dirty_pages(mapping
);
288 put_cpu_var(ratelimits
);
290 EXPORT_SYMBOL(balance_dirty_pages_ratelimited
);
292 void throttle_vm_writeout(void)
294 struct writeback_state wbs
;
295 long background_thresh
;
299 get_dirty_limits(&wbs
, &background_thresh
, &dirty_thresh
, NULL
);
302 * Boost the allowable dirty threshold a bit for page
303 * allocators so they don't get DoS'ed by heavy writers
305 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
307 if (wbs
.nr_unstable
+ wbs
.nr_writeback
<= dirty_thresh
)
309 blk_congestion_wait(WRITE
, HZ
/10);
315 * writeback at least _min_pages, and keep writing until the amount of dirty
316 * memory is less than the background threshold, or until we're all clean.
318 static void background_writeout(unsigned long _min_pages
)
320 long min_pages
= _min_pages
;
321 struct writeback_control wbc
= {
323 .sync_mode
= WB_SYNC_NONE
,
324 .older_than_this
= NULL
,
330 struct writeback_state wbs
;
331 long background_thresh
;
334 get_dirty_limits(&wbs
, &background_thresh
, &dirty_thresh
, NULL
);
335 if (wbs
.nr_dirty
+ wbs
.nr_unstable
< background_thresh
338 wbc
.encountered_congestion
= 0;
339 wbc
.nr_to_write
= MAX_WRITEBACK_PAGES
;
340 wbc
.pages_skipped
= 0;
341 writeback_inodes(&wbc
);
342 min_pages
-= MAX_WRITEBACK_PAGES
- wbc
.nr_to_write
;
343 if (wbc
.nr_to_write
> 0 || wbc
.pages_skipped
> 0) {
344 /* Wrote less than expected */
345 blk_congestion_wait(WRITE
, HZ
/10);
346 if (!wbc
.encountered_congestion
)
353 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
354 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
355 * -1 if all pdflush threads were busy.
357 int wakeup_pdflush(long nr_pages
)
360 struct writeback_state wbs
;
362 get_writeback_state(&wbs
);
363 nr_pages
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
365 return pdflush_operation(background_writeout
, nr_pages
);
368 static void wb_timer_fn(unsigned long unused
);
369 static void laptop_timer_fn(unsigned long unused
);
371 static DEFINE_TIMER(wb_timer
, wb_timer_fn
, 0, 0);
372 static DEFINE_TIMER(laptop_mode_wb_timer
, laptop_timer_fn
, 0, 0);
375 * Periodic writeback of "old" data.
377 * Define "old": the first time one of an inode's pages is dirtied, we mark the
378 * dirtying-time in the inode's address_space. So this periodic writeback code
379 * just walks the superblock inode list, writing back any inodes which are
380 * older than a specific point in time.
382 * Try to run once per dirty_writeback_centisecs. But if a writeback event
383 * takes longer than a dirty_writeback_centisecs interval, then leave a
386 * older_than_this takes precedence over nr_to_write. So we'll only write back
387 * all dirty pages if they are all attached to "old" mappings.
389 static void wb_kupdate(unsigned long arg
)
391 unsigned long oldest_jif
;
392 unsigned long start_jif
;
393 unsigned long next_jif
;
395 struct writeback_state wbs
;
396 struct writeback_control wbc
= {
398 .sync_mode
= WB_SYNC_NONE
,
399 .older_than_this
= &oldest_jif
,
407 get_writeback_state(&wbs
);
408 oldest_jif
= jiffies
- (dirty_expire_centisecs
* HZ
) / 100;
410 next_jif
= start_jif
+ (dirty_writeback_centisecs
* HZ
) / 100;
411 nr_to_write
= wbs
.nr_dirty
+ wbs
.nr_unstable
+
412 (inodes_stat
.nr_inodes
- inodes_stat
.nr_unused
);
413 while (nr_to_write
> 0) {
414 wbc
.encountered_congestion
= 0;
415 wbc
.nr_to_write
= MAX_WRITEBACK_PAGES
;
416 writeback_inodes(&wbc
);
417 if (wbc
.nr_to_write
> 0) {
418 if (wbc
.encountered_congestion
)
419 blk_congestion_wait(WRITE
, HZ
/10);
421 break; /* All the old data is written */
423 nr_to_write
-= MAX_WRITEBACK_PAGES
- wbc
.nr_to_write
;
425 if (time_before(next_jif
, jiffies
+ HZ
))
426 next_jif
= jiffies
+ HZ
;
427 if (dirty_writeback_centisecs
)
428 mod_timer(&wb_timer
, next_jif
);
432 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
434 int dirty_writeback_centisecs_handler(ctl_table
*table
, int write
,
435 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
437 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
438 if (dirty_writeback_centisecs
) {
440 jiffies
+ (dirty_writeback_centisecs
* HZ
) / 100);
442 del_timer(&wb_timer
);
447 static void wb_timer_fn(unsigned long unused
)
449 if (pdflush_operation(wb_kupdate
, 0) < 0)
450 mod_timer(&wb_timer
, jiffies
+ HZ
); /* delay 1 second */
453 static void laptop_flush(unsigned long unused
)
458 static void laptop_timer_fn(unsigned long unused
)
460 pdflush_operation(laptop_flush
, 0);
464 * We've spun up the disk and we're in laptop mode: schedule writeback
465 * of all dirty data a few seconds from now. If the flush is already scheduled
466 * then push it back - the user is still using the disk.
468 void laptop_io_completion(void)
470 mod_timer(&laptop_mode_wb_timer
, jiffies
+ laptop_mode
* HZ
);
474 * We're in laptop mode and we've just synced. The sync's writes will have
475 * caused another writeback to be scheduled by laptop_io_completion.
476 * Nothing needs to be written back anymore, so we unschedule the writeback.
478 void laptop_sync_completion(void)
480 del_timer(&laptop_mode_wb_timer
);
484 * If ratelimit_pages is too high then we can get into dirty-data overload
485 * if a large number of processes all perform writes at the same time.
486 * If it is too low then SMP machines will call the (expensive)
487 * get_writeback_state too often.
489 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
490 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
491 * thresholds before writeback cuts in.
493 * But the limit should not be set too high. Because it also controls the
494 * amount of memory which the balance_dirty_pages() caller has to write back.
495 * If this is too large then the caller will block on the IO queue all the
496 * time. So limit it to four megabytes - the balance_dirty_pages() caller
497 * will write six megabyte chunks, max.
500 static void set_ratelimit(void)
502 ratelimit_pages
= total_pages
/ (num_online_cpus() * 32);
503 if (ratelimit_pages
< 16)
504 ratelimit_pages
= 16;
505 if (ratelimit_pages
* PAGE_CACHE_SIZE
> 4096 * 1024)
506 ratelimit_pages
= (4096 * 1024) / PAGE_CACHE_SIZE
;
510 ratelimit_handler(struct notifier_block
*self
, unsigned long u
, void *v
)
516 static struct notifier_block ratelimit_nb
= {
517 .notifier_call
= ratelimit_handler
,
522 * If the machine has a large highmem:lowmem ratio then scale back the default
523 * dirty memory thresholds: allowing too much dirty highmem pins an excessive
524 * number of buffer_heads.
526 void __init
page_writeback_init(void)
528 long buffer_pages
= nr_free_buffer_pages();
531 total_pages
= nr_free_pagecache_pages();
533 correction
= (100 * 4 * buffer_pages
) / total_pages
;
535 if (correction
< 100) {
536 dirty_background_ratio
*= correction
;
537 dirty_background_ratio
/= 100;
538 vm_dirty_ratio
*= correction
;
539 vm_dirty_ratio
/= 100;
541 if (dirty_background_ratio
<= 0)
542 dirty_background_ratio
= 1;
543 if (vm_dirty_ratio
<= 0)
546 mod_timer(&wb_timer
, jiffies
+ (dirty_writeback_centisecs
* HZ
) / 100);
548 register_cpu_notifier(&ratelimit_nb
);
551 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
553 if (wbc
->nr_to_write
<= 0)
555 if (mapping
->a_ops
->writepages
)
556 return mapping
->a_ops
->writepages(mapping
, wbc
);
557 return generic_writepages(mapping
, wbc
);
561 * write_one_page - write out a single page and optionally wait on I/O
563 * @page: the page to write
564 * @wait: if true, wait on writeout
566 * The page must be locked by the caller and will be unlocked upon return.
568 * write_one_page() returns a negative error code if I/O failed.
570 int write_one_page(struct page
*page
, int wait
)
572 struct address_space
*mapping
= page
->mapping
;
574 struct writeback_control wbc
= {
575 .sync_mode
= WB_SYNC_ALL
,
579 BUG_ON(!PageLocked(page
));
582 wait_on_page_writeback(page
);
584 if (clear_page_dirty_for_io(page
)) {
585 page_cache_get(page
);
586 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
587 if (ret
== 0 && wait
) {
588 wait_on_page_writeback(page
);
592 page_cache_release(page
);
598 EXPORT_SYMBOL(write_one_page
);
601 * For address_spaces which do not use buffers. Just tag the page as dirty in
604 * This is also used when a single buffer is being dirtied: we want to set the
605 * page dirty in that case, but not all the buffers. This is a "bottom-up"
606 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
608 * Most callers have locked the page, which pins the address_space in memory.
609 * But zap_pte_range() does not lock the page, however in that case the
610 * mapping is pinned by the vma's ->vm_file reference.
612 * We take care to handle the case where the page was truncated from the
613 * mapping by re-checking page_mapping() insode tree_lock.
615 int __set_page_dirty_nobuffers(struct page
*page
)
619 if (!TestSetPageDirty(page
)) {
620 struct address_space
*mapping
= page_mapping(page
);
621 struct address_space
*mapping2
;
624 write_lock_irq(&mapping
->tree_lock
);
625 mapping2
= page_mapping(page
);
626 if (mapping2
) { /* Race with truncate? */
627 BUG_ON(mapping2
!= mapping
);
628 if (mapping_cap_account_dirty(mapping
))
629 inc_page_state(nr_dirty
);
630 radix_tree_tag_set(&mapping
->page_tree
,
631 page_index(page
), PAGECACHE_TAG_DIRTY
);
633 write_unlock_irq(&mapping
->tree_lock
);
635 /* !PageAnon && !swapper_space */
636 __mark_inode_dirty(mapping
->host
,
643 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
646 * When a writepage implementation decides that it doesn't want to write this
647 * page for some reason, it should redirty the locked page via
648 * redirty_page_for_writepage() and it should then unlock the page and return 0
650 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
652 wbc
->pages_skipped
++;
653 return __set_page_dirty_nobuffers(page
);
655 EXPORT_SYMBOL(redirty_page_for_writepage
);
658 * If the mapping doesn't provide a set_page_dirty a_op, then
659 * just fall through and assume that it wants buffer_heads.
661 int fastcall
set_page_dirty(struct page
*page
)
663 struct address_space
*mapping
= page_mapping(page
);
665 if (likely(mapping
)) {
666 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
669 return __set_page_dirty_buffers(page
);
671 if (!PageDirty(page
))
675 EXPORT_SYMBOL(set_page_dirty
);
678 * set_page_dirty() is racy if the caller has no reference against
679 * page->mapping->host, and if the page is unlocked. This is because another
680 * CPU could truncate the page off the mapping and then free the mapping.
682 * Usually, the page _is_ locked, or the caller is a user-space process which
683 * holds a reference on the inode by having an open file.
685 * In other cases, the page should be locked before running set_page_dirty().
687 int set_page_dirty_lock(struct page
*page
)
692 ret
= set_page_dirty(page
);
696 EXPORT_SYMBOL(set_page_dirty_lock
);
699 * Clear a page's dirty flag, while caring for dirty memory accounting.
700 * Returns true if the page was previously dirty.
702 int test_clear_page_dirty(struct page
*page
)
704 struct address_space
*mapping
= page_mapping(page
);
708 write_lock_irqsave(&mapping
->tree_lock
, flags
);
709 if (TestClearPageDirty(page
)) {
710 radix_tree_tag_clear(&mapping
->page_tree
,
712 PAGECACHE_TAG_DIRTY
);
713 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
714 if (mapping_cap_account_dirty(mapping
))
715 dec_page_state(nr_dirty
);
718 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
721 return TestClearPageDirty(page
);
723 EXPORT_SYMBOL(test_clear_page_dirty
);
726 * Clear a page's dirty flag, while caring for dirty memory accounting.
727 * Returns true if the page was previously dirty.
729 * This is for preparing to put the page under writeout. We leave the page
730 * tagged as dirty in the radix tree so that a concurrent write-for-sync
731 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
732 * implementation will run either set_page_writeback() or set_page_dirty(),
733 * at which stage we bring the page's dirty flag and radix-tree dirty tag
736 * This incoherency between the page's dirty flag and radix-tree tag is
737 * unfortunate, but it only exists while the page is locked.
739 int clear_page_dirty_for_io(struct page
*page
)
741 struct address_space
*mapping
= page_mapping(page
);
744 if (TestClearPageDirty(page
)) {
745 if (mapping_cap_account_dirty(mapping
))
746 dec_page_state(nr_dirty
);
751 return TestClearPageDirty(page
);
754 int test_clear_page_writeback(struct page
*page
)
756 struct address_space
*mapping
= page_mapping(page
);
762 write_lock_irqsave(&mapping
->tree_lock
, flags
);
763 ret
= TestClearPageWriteback(page
);
765 radix_tree_tag_clear(&mapping
->page_tree
,
767 PAGECACHE_TAG_WRITEBACK
);
768 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
770 ret
= TestClearPageWriteback(page
);
775 int test_set_page_writeback(struct page
*page
)
777 struct address_space
*mapping
= page_mapping(page
);
783 write_lock_irqsave(&mapping
->tree_lock
, flags
);
784 ret
= TestSetPageWriteback(page
);
786 radix_tree_tag_set(&mapping
->page_tree
,
788 PAGECACHE_TAG_WRITEBACK
);
789 if (!PageDirty(page
))
790 radix_tree_tag_clear(&mapping
->page_tree
,
792 PAGECACHE_TAG_DIRTY
);
793 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
795 ret
= TestSetPageWriteback(page
);
800 EXPORT_SYMBOL(test_set_page_writeback
);
803 * Return true if any of the pages in the mapping are marged with the
806 int mapping_tagged(struct address_space
*mapping
, int tag
)
811 read_lock_irqsave(&mapping
->tree_lock
, flags
);
812 ret
= radix_tree_tagged(&mapping
->page_tree
, tag
);
813 read_unlock_irqrestore(&mapping
->tree_lock
, flags
);
816 EXPORT_SYMBOL(mapping_tagged
);